Making harm visible: Survivor-centred analysis of domestic homicide reviews using AI

Domestic homicide reviews (DHRs) are official reports in the UK written after a person is killed by a partner or family member to understand what went wrong and what actions could prevent future deaths. They should turn a fatal abuse case into lessons for prevention. However, these reviews are often lengthy, time and resource-intensive, and result in vague or broad agency recommendations, usually at the local level.

Police forces and review panels are now exploring how to use AI large language models for these reviews, but it is unclear whether this digital turn will enhance or erode equity, accountability and victim voice, which should be central to domestic homicide review processes.

This project investigates how AI large language models are and will be used to write and analyse domestic homicide reviews, improving transparency, equity and organisational learning. Specifically, the project will work with survivors to scrutinise whether, and under what conditions, large language models can be built and evaluated as a ‘digital good’ when drafting and analysing domestic homicide reviews.

Findings will inform strategies for the responsible deployment of AI to support:

• the analysis of domestic homicide reviews, improving the recognition of harm that may be more hidden, such as in cases of coercive control
• the writing of DHRs, helping agencies, review boards and courts produce more timely, actionable and survivor-centred recommendations

Aims and objectives

The project aims to transform how DHRs are produced and analysed in the UK by addressing delays, inconsistencies and gaps in equity and survivor voice. Objectives include:

• building a large-scale, harmonised dataset of around 1,300 DHRs
• co-producing a survivor-informed codebook of harms and equity factors
• developing and evaluating an explainable AI pipeline for extracting key entities, timelines and recommendations
• comparing AI-generated recommendations with human-authored ones to assess quality and bias
• translating findings into practical guidance for responsible AI use in DHRs

Data analysis

The project will use a mixed-method approach, combining computational techniques with qualitative, survivor-informed evaluation to ensure both technical rigor and ethical accountability. We recognise the need to be transparent so that reviewers can understand the steps taken before working with the Voice of Survivors lived experience panel.

This study involves domestic abuse professionals and the Voice of Survivors lived experience group who have been working with the principal investigator across various research activities since 2022, plus survivors supported by domestic abuse services.